Final Report Summary - NEUCOD (Information-theoretic approach to infer encoding patterns in a decision making process)
Project objectives
In this project we have taken a new approach to unravel the information encoding and communication mechanisms underlying a cognitive task. The whole project is divided into three main scientific tasks:
1) Task 1: To propose a neural coding inference method based on neuronal correlations that explains how sensory and behavioral information is distributedly encoded by neuron pairs (in a first stage) as well as larger ensembles of neurons (second stage).
2)Task 2: To identify encoding patterns in a simulated model as well as in simultaneously recorded data from monkeys performing a decision making task using methods of Task 1.
3)Task 3: To associate population coding with different degrees of brain cognitive disability.
Description of the work, main results and expected impact
Since the beginning of the project we worked across the three main tasks without strictly following a sequential implementation. This reorganization was due to two facts:
_The opportunity and demand to analyze a rich dataset of cortical simultaneous recordings from two monkeys (rather than one) while performing a decision making task.
_ A new collaboration that we built with members of the Epilepsy Unit at “Hospital del Mar”, Barcelona, Spain, which allowed us to have access to human intracranial recordings from epileptic patients and tackle the objectives of Task 3 at a human level.
As a consequence, we first implemented Task 1 and simultaneously applied its outcomes to monkey data in Task 2 and human data in Task 3.
Next, we will give further details on the progress and achievements in each task.
1) Neural-coding inference method: We proposed a simple method to analyze how information may be jointly encoded (or distributed) by neurons across cortical areas engaged in a cognitive task. This methodology combines
_a statistical model of neuronal spike trains using Markovian processes
_a directional information-theoretic measure over pairs of discrete sequences.
_a statistical test on the dependence of this measure on different task variables (stimulus frequency, type of response, etc.) for different time delays.
By using this method, one can detect those neuron pairs that jointly encode information about stimuli or responses and relate these “informative links” to neurons that individually encode information. As a result, one can track how stimuli information that is encoded at a particular sensory neuron is later distributed to other neurons. When the encoded variable is binary (such as the decision reported by our two monkeys), we can categorize the different encoding patterns in three sets, “ON-ON” set if the neuron pair is significantly (in a statistical sense) correlated for each decision report, and “ON-OFF” and “OFF-ON” if the neuron pair is significantly correlated for each decision report respectively. Moreover, the method also identifies the delays at which these significant correlations occur. Interestingly, in the two monkeys that we analyzed most encoding patterns are in practice of the type “ON-OFF” and “OFF-ON”, and hence, this could suggest a neural mechanism by which neuronal correlation pathways become activated for specific stimulus features or decision reports.
The proposed method is a priori not sensitive to non-stationarities in the data and relies on a pre-binning of the data recordings and a division of the whole task timeline into distinct periods. Hence, we are working towards making the method time-continuous and sensitive to changes in the data statistics.
Overall, this method offers a systematic method to extract patterns of distributed information from simultaneous recordings, which has been proved to yield consistent results with the existing literature. As the technology to simultaneously record single neurons progresses rapidly (in animals as well as humans), this method can have a huge impact in further scientific studies as a computational tool to infer how information on specific variables is shared and transferred across recorded channel activity.
The details of these results were published in the supplementary information of the article “Task-driven intra- and inter-area communications in primate cerebral cortex”, Proceedings of National Academy of Sciences (PNAS), vol. 112, pp. 4761–4766, 2015.
2) Identification of encoding patterns in real data: We have analyzed cortical simultaneous recordings of two monkeys performing a discrimination task along the lines described above. Our most interesting results from a biological perspective are:
_The neural coding mechanisms to encode and distribute information by the monkey are task-specific, that is encoding patterns that emerge across cortical areas during the original task are no longer active during a control task, in which the monkey has no incentive to perform the task.
_Stimuli and decision reports are mostly encoded in the patterns “ON-OFF” and “OFF-ON” suggesting that information on distinct stimulus or response features is encoded in disjoint correlation pathways.
_The delays associated encoding patterns of stimulus and responses are significantly different suggesting that sensory and behavioral information are distributed at different timescales.
We are now working in applying this methodology to predict correct and error trials based on sensory and behavioral encoding patterns at the first stages of the task. Additionally, we are also quantifying information transfer between thalamus and cortex of one money performing a detection task. The potential impact of these type of findings is important as they may provide a mechanistic model on how neural information is processed and communicated across neurons.
The details of these results were published in the main article of “Task-driven intra- and inter-area communications in primate cerebral cortex” Proceedings of National Academy of Sciences (PNAS), vol. 112, pp. 4761–4766, 2015 and will be published as a book chapter in “The neurotechnology revolution: new technologies for understanding and manipulating brain circuits and computations”, Wiley Editions, 2016.
3) Functional interpretation of population encoding patterns. We have started to analyze intracranial EEG recordings from epileptic patients to develop a multivariate analysis that infers connectivity states of the implanted/recorded area and track how these states evolve over time during spontaneous activity as well as task-driven activity. During spontaneous activity, these states may provide information about the temporal occurrence of seizures and may predict the clinical outcome of patients. In addition, we can use this analytical method on the same recordings while the patient is performing a decision-making task. Specifically, during task performance we can evaluate neural encoding patterns during correct and incorrect responses and disentangle its different encoding mechanisms.
All in all, this method may help identify neural correlates of epileptic cognitive deficit in humans. In particular, the state detection algorithm may improve seizure prediction and help neurosurgeons better localize the epileptic focus that they need to resect at the end of the monitoring stage. Preliminary results on this task were presented at the 2014 Society for Neuroscience conference (November 15-19 2014 in Washington DC, US), the 2015’ International Workshop on Seizure Prediction (August 2-6 2015 in Melbourne, Australia, poster attached). Two students currently work along these research topics.
In this project we have taken a new approach to unravel the information encoding and communication mechanisms underlying a cognitive task. The whole project is divided into three main scientific tasks:
1) Task 1: To propose a neural coding inference method based on neuronal correlations that explains how sensory and behavioral information is distributedly encoded by neuron pairs (in a first stage) as well as larger ensembles of neurons (second stage).
2)Task 2: To identify encoding patterns in a simulated model as well as in simultaneously recorded data from monkeys performing a decision making task using methods of Task 1.
3)Task 3: To associate population coding with different degrees of brain cognitive disability.
Description of the work, main results and expected impact
Since the beginning of the project we worked across the three main tasks without strictly following a sequential implementation. This reorganization was due to two facts:
_The opportunity and demand to analyze a rich dataset of cortical simultaneous recordings from two monkeys (rather than one) while performing a decision making task.
_ A new collaboration that we built with members of the Epilepsy Unit at “Hospital del Mar”, Barcelona, Spain, which allowed us to have access to human intracranial recordings from epileptic patients and tackle the objectives of Task 3 at a human level.
As a consequence, we first implemented Task 1 and simultaneously applied its outcomes to monkey data in Task 2 and human data in Task 3.
Next, we will give further details on the progress and achievements in each task.
1) Neural-coding inference method: We proposed a simple method to analyze how information may be jointly encoded (or distributed) by neurons across cortical areas engaged in a cognitive task. This methodology combines
_a statistical model of neuronal spike trains using Markovian processes
_a directional information-theoretic measure over pairs of discrete sequences.
_a statistical test on the dependence of this measure on different task variables (stimulus frequency, type of response, etc.) for different time delays.
By using this method, one can detect those neuron pairs that jointly encode information about stimuli or responses and relate these “informative links” to neurons that individually encode information. As a result, one can track how stimuli information that is encoded at a particular sensory neuron is later distributed to other neurons. When the encoded variable is binary (such as the decision reported by our two monkeys), we can categorize the different encoding patterns in three sets, “ON-ON” set if the neuron pair is significantly (in a statistical sense) correlated for each decision report, and “ON-OFF” and “OFF-ON” if the neuron pair is significantly correlated for each decision report respectively. Moreover, the method also identifies the delays at which these significant correlations occur. Interestingly, in the two monkeys that we analyzed most encoding patterns are in practice of the type “ON-OFF” and “OFF-ON”, and hence, this could suggest a neural mechanism by which neuronal correlation pathways become activated for specific stimulus features or decision reports.
The proposed method is a priori not sensitive to non-stationarities in the data and relies on a pre-binning of the data recordings and a division of the whole task timeline into distinct periods. Hence, we are working towards making the method time-continuous and sensitive to changes in the data statistics.
Overall, this method offers a systematic method to extract patterns of distributed information from simultaneous recordings, which has been proved to yield consistent results with the existing literature. As the technology to simultaneously record single neurons progresses rapidly (in animals as well as humans), this method can have a huge impact in further scientific studies as a computational tool to infer how information on specific variables is shared and transferred across recorded channel activity.
The details of these results were published in the supplementary information of the article “Task-driven intra- and inter-area communications in primate cerebral cortex”, Proceedings of National Academy of Sciences (PNAS), vol. 112, pp. 4761–4766, 2015.
2) Identification of encoding patterns in real data: We have analyzed cortical simultaneous recordings of two monkeys performing a discrimination task along the lines described above. Our most interesting results from a biological perspective are:
_The neural coding mechanisms to encode and distribute information by the monkey are task-specific, that is encoding patterns that emerge across cortical areas during the original task are no longer active during a control task, in which the monkey has no incentive to perform the task.
_Stimuli and decision reports are mostly encoded in the patterns “ON-OFF” and “OFF-ON” suggesting that information on distinct stimulus or response features is encoded in disjoint correlation pathways.
_The delays associated encoding patterns of stimulus and responses are significantly different suggesting that sensory and behavioral information are distributed at different timescales.
We are now working in applying this methodology to predict correct and error trials based on sensory and behavioral encoding patterns at the first stages of the task. Additionally, we are also quantifying information transfer between thalamus and cortex of one money performing a detection task. The potential impact of these type of findings is important as they may provide a mechanistic model on how neural information is processed and communicated across neurons.
The details of these results were published in the main article of “Task-driven intra- and inter-area communications in primate cerebral cortex” Proceedings of National Academy of Sciences (PNAS), vol. 112, pp. 4761–4766, 2015 and will be published as a book chapter in “The neurotechnology revolution: new technologies for understanding and manipulating brain circuits and computations”, Wiley Editions, 2016.
3) Functional interpretation of population encoding patterns. We have started to analyze intracranial EEG recordings from epileptic patients to develop a multivariate analysis that infers connectivity states of the implanted/recorded area and track how these states evolve over time during spontaneous activity as well as task-driven activity. During spontaneous activity, these states may provide information about the temporal occurrence of seizures and may predict the clinical outcome of patients. In addition, we can use this analytical method on the same recordings while the patient is performing a decision-making task. Specifically, during task performance we can evaluate neural encoding patterns during correct and incorrect responses and disentangle its different encoding mechanisms.
All in all, this method may help identify neural correlates of epileptic cognitive deficit in humans. In particular, the state detection algorithm may improve seizure prediction and help neurosurgeons better localize the epileptic focus that they need to resect at the end of the monitoring stage. Preliminary results on this task were presented at the 2014 Society for Neuroscience conference (November 15-19 2014 in Washington DC, US), the 2015’ International Workshop on Seizure Prediction (August 2-6 2015 in Melbourne, Australia, poster attached). Two students currently work along these research topics.
